Amazon / Safety at scale

Safety has to survive production.

I built and scaled AI safety systems, led cross-team delivery, and made regional launches faster and more repeatable.

AWS · Senior Software Development Engineer
Nov 2023–Mar 2026

Enterprise AI safetyFrom architecture to customer adoption.

Safety engineeringInference operationsCross-team delivery

Connecting system design, integration, and operational readiness.

The challenge

Protection, latency, and delivery all had to work together.

Safety checks sat on customer-facing AI workflows. Adding coverage increased coordination across services and teams, while every extra call affected latency and cost. My work spanned the processing layer, inference fleet, customer integration, and rollout.

60–70%

Lower safety latency

50%

Lower safety API cost

50%

Smaller inference fleet

Engineering decisions

Coverage, efficiency, and repeatability.

Explore the work across safety, inference, and delivery.

01 / Safety

Build security collaboration into delivery.

I partnered with the AppSec red team to harden the service against common LLM attack classes. My work connected safety engineering with production readiness across multi-modal workflows.

The roadmap connected cost, latency, compliance, and technical debt across safety workflows.

02 / Latency

Remove waiting and redundant work.

Concurrent checks, direct payload support, and model/provider migration cut safety latency by 60–70%. Removing redundant checks reduced safety API cost by 50%.

Parallel processing helped add safety coverage while limiting latency overhead.

03 / Capacity

Size the fleet from load-test evidence.

I operated a multi-region GPU inference fleet. Load-test-based sizing and target-tracking autoscaling reduced fleet size by 50% while maintaining service availability.

Per-modality dashboards tracked tail latency, memory, and thread utilization to support capacity decisions.

04 / Delivery

Make a new region a repeatable launch.

I led regional launch automation across infrastructure and operational readiness. The work connected quotas, canaries, alarms, dashboards, and compliance.

Reusable automation accelerated regional launches and reduced manual coordination.

The coordination challenge

Multiple teams had to ship a coherent customer experience.

I led 8 engineers across 5 teams to deliver enterprise PII detection and redaction in 18 weeks. We consolidated distributed components behind a unified safety-processing API across 4 modalities, simplifying customer integration.

I worked directly with 4+ enterprise customers and solutions architects, from early solution design through beta launch. That work connected integration constraints to the API and rollout decisions.

Regional delivery

Make readiness repeatable.

Reusable infrastructure and operational workflows cut launch time across 5+ regions.

Before

1–2 months

After

Under 2 weeks across 5+ regions

Beyond the launch

Build the team’s ability to operate the system.

I supported team growth through technical interviews and mentorship, and wrote operating procedures for on-call, incident response, patching, and operational reviews.

Alongside the safety roadmap, I shortened local test cycles and automated manual release validation, reducing repetitive engineering work.

In practice

60–70% lower safety latency.
50% lower safety API cost and fleet size.

Delivered alongside enterprise PII coverage, sustained service availability, and faster regional launches.